langgraph-checkpoint-upstash-redis 0.1.0

LangGraph checkpoint saver backed by Upstash Redis (REST API)

langgraph-checkpoint-upstash-redis

A LangGraph checkpoint saver backed by Upstash Redis over its REST API.

Why this exists

The official langgraph-checkpoint-redis package relies on RediSearch and RedisJSON to index checkpoints. Upstash Redis does not support those modules, so that package doesn't work against an Upstash Redis database.

This package sidesteps that entirely: checkpoints are stored as plain Redis hashes and indexed with sorted sets, so it only needs commands every Redis-compatible server (including Upstash) supports. It talks to Upstash exclusively over REST via the upstash-redis SDK, which makes it a good fit for serverless Python (AWS Lambda, Vercel, Cloud Run, etc.) where long-lived TCP connections are awkward or unavailable.

Install

pip install langgraph-checkpoint-upstash-redis

Usage

Sync

from langgraph.checkpoint.upstash_redis import UpstashRedisSaver

# Reads UPSTASH_REDIS_REST_URL / UPSTASH_REDIS_REST_TOKEN from the environment.
saver = UpstashRedisSaver.from_env()

# Or pass credentials explicitly:
saver = UpstashRedisSaver.from_conn_info(url="...", token="...")

# use checkpoint in an agent
from langchain.agents import create_agent

agent = create_agent(model="groq:openai/gpt-oss-120b", checkpointer=saver)

thread_config = {"configurable": {"thread_id": "1"}}
response = agent.invoke(
    {"messages": [{"role": "user", "content": "what is my name?"}]},
    thread_config,
)["messages"][-1].content

print(response) # Hello, Bob! Nice to meet you. How can I help you today?

response = agent.invoke(
    {"messages": [{"role": "user", "content": "what is my name?"}]},
    thread_config,
)["messages"][-1].content

print(response) # You mentioned that your name is **Bob**.

Async

from langgraph.checkpoint.upstash_redis.aio import AsyncUpstashRedisSaver

# use checkpoint in an agent
from langchain.agents import create_agent

agent = create_agent(model="groq:openai/gpt-oss-120b", checkpointer=saver)

thread_config = {"configurable": {"thread_id": "1"}}
response = await agent.ainvoke(
    {"messages": [{"role": "user", "content": "what is my name?"}]},
    thread_config,
)["messages"][-1].content

print(response) # Hello, Bob! Nice to meet you. How can I help you today?

response = await agent.ainvoke(
    {"messages": [{"role": "user", "content": "what is my name?"}]},
    thread_config,
)["messages"][-1].content

print(response) # You mentioned that your name is **Bob**.

Use AsyncUpstashRedisSaver with graph.ainvoke/astream/etc.; use UpstashRedisSaver for sync graph execution. Each saver only implements the sync or async half of the BaseCheckpointSaver interface, matching the pattern used by other single-mode checkpointers (e.g. AsyncSqliteSaver).

TTL

Both savers accept an optional ttl (seconds). When set, checkpoints and their indices expire automatically:

saver = UpstashRedisSaver.from_env(ttl=60 * 60 * 24)  # 1 day

Design notes

• No RediSearch / RedisJSON. Checkpoints are stored as Redis hashes (checkpoint:{thread_id}:{checkpoint_ns}:{checkpoint_id}) and indexed per thread/namespace with a sorted set (checkpoints_idx:{thread_id}:{checkpoint_ns}). LangGraph's checkpoint IDs are monotonically increasing strings, so lexicographic order (ZRANGEBYLEX) is equivalent to chronological order — no search index needed.
• REST-only. All I/O goes through the upstash-redis HTTP client. Writes to a checkpoint (the hash, its index entry, and namespace/thread bookkeeping sets) are batched into a single pipelined HTTP request.
• Metadata filtering (the filter argument to list) is applied client-side after fetching candidate checkpoints for the relevant thread/namespace, since there's no secondary index to push it down to. This mirrors what LangGraph's own InMemorySaver does and is fine at the checkpoint volumes a single thread accumulates.
• Binary checkpoint/metadata/write payloads (from LangGraph's JsonPlusSerializer) are base64-encoded before being stored in Redis hash fields, since the REST protocol is text-based.

Out of scope for this package: a BaseStore implementation, vector search, and the DeltaChannel / prune / copy_thread / delete_for_runs checkpointer extensions — only the core checkpoint CRUD surface is implemented.

Testing

Unit tests for serialization and key-building run without any network access:

pip install -e ".[dev]"
pytest tests/test_base.py

Integration tests exercise a real Upstash Redis database and require UPSTASH_REDIS_REST_URL / UPSTASH_REDIS_REST_TOKEN to be set (a free Upstash database works fine); they're skipped automatically otherwise:

export UPSTASH_REDIS_REST_URL=...
export UPSTASH_REDIS_REST_TOKEN=...
pytest tests/

License

MIT